A Dutch startup wants $100 million to build a chip that it says makes Nvidia's best inference hardware look like a Ford Pinto trying to race a Tesla. The chip is called CRAFTWERK. The company is Euclyd. And the people backing it include the former CEO of ASML, the co-inventor of the Intel 4004, and the founder of Elastic.
The specs Euclyd has published are not small. Each CRAFTWERK system-in-package integrates 16,384 custom SIMD processors alongside 1 terabyte of custom ultra-bandwidth memory, feeding up to 32 petaflops of FP4 compute. Rack-scale systems stack 32 of these packages together for a claimed 1.024 exaflops of FP4 performance at 125 kilowatts. Euclyd says this represents a 100x improvement in tokens per joule compared to current leading inference hardware, based on modeling against Llama 4 Maverick EE News Europe.
These numbers are not independently verified. There are no third-party benchmarks, no peer-reviewed results, and no shipped silicon. The production target is 2028. Euclyd was founded in 2024 and has a seed round of less than 10 million euros Ioplus. The gap between architecture announcement and working silicon is where most chip startups die.
But the interesting story is not whether Euclyd delivers. It is what happens to several hundred billion dollars of sovereign infrastructure planning if any company in this category does.
Gulf sovereign wealth funds collectively deployed $66 billion into AI and digitization in 2025 TechWeez. Abu Dhabi's MGX is targeting $100 billion or more in AI-related assets under management, with Stargate UAE — a hyperscale data center campus involving G42, OpenAI, Oracle, and SoftBank — as its flagship project TechWeez. Saudi Arabia's Public Investment Fund launched Project Transcendence, a $100 billion AI infrastructure initiative, in 2024, followed by HUMAIN in 2025 TechWeez. France announced a €109 billion AI investment plan as part of a broader European effort Introl. MENA data center capacity is projected to triple from 1 gigawatt to 3.3 gigawatts by 2030 TechWeez.
Every one of these investments was sized around the assumption that compute remains expensive, that bandwidth between processor and memory stays a fundamental constraint, and that Nvidia's hardware defines the cost ceiling for AI inference at scale. That assumption is now being tested by a category of startups claiming a different physics.
The architecture Euclyd is pursuing abandons the prevailing model. Current GPU-based inference systems spend significant energy shuttling data between separate processor and memory dies. Euclyd's approach processes information in multiple places simultaneously, keeping compute and memory tightly integrated Ioplus. This is not a new idea — in-memory computing projects have existed for decades and consistently ran into the same wall: yield problems at scale, toolchain gaps, and the difficulty of building a software ecosystem around non-standard architectures.
Euclyd's argument is that the economics of AI inference at hyperscale have finally made the tradeoff worth taking seriously. Patrick Schneider-Sikorsky of the NATO Innovation Fund put it plainly in an interview: the current GPU architecture was simply not built for the inference scale now required Ioplus. That is a structural observation, not a company-specific claim. And it is being made at the same moment that sovereign funds are committing to build out GPU-dependent infrastructure on a multi-decade timescale.
Nvidia is not standing still. It has invested $4 billion in photonic computing companies in the past year Ioplus, a signal that even the dominant player sees the memory-bandwidth wall as a real constraint rather than a solved problem.
The uncomfortable question for sovereign AI infrastructure planners is not whether Euclyd specifically will succeed. It is whether the entire category of energy-efficient inference architecture — if any player in it reaches production — would require a fundamental revision of the capex models that underpin current data center planning. A 100x improvement in tokens per joule does not just change a company's unit economics. It changes the answer to the question every sovereign fund has already answered: how much does inference cost, and how much should we spend to own it.
Euclyd is raising $100 million to find out if its answer is right. The people who already bet on it include some of the most credible names in semiconductor history — former ASML CEO Peter Wennink, microprocessor co-inventor Federico Faggin, and Elastic founder Steven Schuurman EE News Europe. Whether that is a signal of genuine technical confidence or the prestige premium that attaches to famous names in a hot sector is exactly the question that a due diligence process — and eventually, silicon — will answer.
